Machine Learning for Streaming Data with Python (Joos Korstanje) (z-library.sk, 1lib.sk, z-lib.sk)

【One-Line Pitch】 A hands-on guide to building real-time machine learning systems with Python, showing how to move beyond batch thinking toward continuously learning models. Best suited for data scientists and ML practitioners who already know standard scikit-learn workflows and now need to handle live, unbounded data streams. 【Book Arc】 - **Opening (~0%–15%)**: Frames the core problem — most analytics are still done in batch even when real-time value is possible — and contrasts batch versus streaming approaches, including the advantages of real-time data-generating processes. - **Early (~15%–30%)**: Builds the architectural foundation: turning analytics into functions, sending HTTP requests, and standing up a minimal real-time inference API (e.g., via AWS API Gateway and Lambda), while flagging the software-engineering concerns a data scientist should not own alone. - **Early–Middle (~30%–45%)**: Covers streaming descriptive statistics (rolling windows, mode, standard deviation), real-time visualization with Plotly Dash, and alerting/monitoring systems, including statistical process control for more advanced alerts. - **Middle (~45%–55%)**: Introduces online machine learning as distinct from offline learning — models that update sequentially rather than in a single training pass — demonstrated with River on the iris dataset and learning-curve inspection. - **Late (~55%–85%)**: Moves into industrial use cases such as online anomaly detection and a PCA alternative that works on real-time data, plus the challenges that arise when streaming and how to mitigate them. - **Ending (~85%–100%)**: Consolidates best practices for generating real-time insights and building an API for real-time machine learning inference, aiming to leave the reader confident deploying streaming models. 【Key Takeaways】 - **Streaming is an architectural choice, not just a modeling one** (Early): the book treats real-time analytics as a pipeline — function, API, deployment — before touching model internals. - **Batch and streaming coexist** (Early): many use cases still run in batch simply because those methods are more familiar; the book helps you judge when streaming adds genuine value. - **Descriptive statistics need rethinking on streams** (Early–Middle): rolling windows, mode, and standard deviation are recomputed over sliding windows rather than fixed datasets. - **Alerting evolves from static rules to statistical process control** (Middle): simple threshold alerts on single values give way to logic based on means, medians, and process stability. - **Online learning updates models sequentially** (Middle): unlike offline train-deploy-retrain cycles, online models keep learning from each new observation, which suits drifting data. - **River is the primary toolkit** (Middle): the book uses River for online models (e.g., logistic regression on iris) and inspects learning curves to judge stability. - **Real-time visualization is achievable with lightweight tools** (Middle): Plotly Dash with interval callbacks provides near-real-time dashboards without heavy infrastructure. - **Deployment and inference APIs close the loop** (Late): the book walks through exposing models for real-time inference, connecting analytics to production consumers. 【Reading Tips】 - Deep-read the early architecture chapters if you have never built an API; skim them if you already deploy services regularly. - Treat the River code examples as the core — type them out rather than copying, since the sequential `learn_one` pattern is easy to misapply. - Pay attention to the "challenges and mitigations" material in the later chapters; that is where the practical judgment lives. - Use the GitHub repository alongside the text; the book explicitly points to notebooks and a real-time visualization script there. - If you are strong on statistics, skim the descriptive-statistics refresher and focus on how windows change the computation. 【Coverage Limits】 The excerpts cover the book's framing, architecture, descriptive statistics, visualization, alerting, and the introduction to online learning with River, but do not detail the specific anomaly detection methods, the PCA alternative implementation, or the full best-practices chapter. Chapter titles and exact percentages beyond the markers shown are not verifiable from the excerpts.